A referenceless light field image quality assessment method and system based on light field decoupling

By using a light field decoupling-based method to screen high-quality data and extracting feature information using a four-stream convolutional neural network, the inadequacy of objective evaluation in light field image quality assessment is addressed, and a highly accurate light field image quality assessment is achieved.

CN117495841BActive Publication Date: 2026-11-17ANHUI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311633609.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-01
Publication Date
2026-11-17
Estimated Expiration
2043-12-01

AI Technical Summary

Technical Problem

Existing methods for evaluating the quality of light field images lack an objective, reference-free evaluation system that aligns with human subjective perception. In particular, objective, reference-free evaluation methods have shortcomings in practical applications.

Method used

A light field decoupling-based approach is adopted, which uses an image patch selection strategy to filter high-quality data, constructs a sample dataset, and uses a four-stream convolutional neural network to extract spatial and angular feature information, and fuses the feature information to predict the light field image quality score.

Benefits of technology

It achieves high-accuracy light field image quality assessment, solves the dimensionality reduction problem of high-dimensional light field image data, improves the model's ability to extract multi-class feature information, and provides accurate objective quality assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117495841B_ABST
    Figure CN117495841B_ABST
Patent Text Reader

Abstract

This invention relates to the field of light field image quality assessment. It discloses a method and system for assessing the quality of no-reference light field images based on light field decoupling. The method includes: using a specially designed image patch selection strategy to filter and reconstruct information from the original light field image to obtain high-quality training samples; using a light field decoupling mechanism on the training samples to transform high-dimensional light field image data into low-dimensional input data for a neural network; designing a multi-stream convolutional neural network to extract features from the input data, learning spatial quality information, angular quality information, and spatial angular consistency information in the light field image; and obtaining a quality prediction value based on the fused feature information. This invention utilizes the decoupling concept combined with a multi-stream convolutional neural network to extract quality-related features of the light field image, resulting in a quality assessment of the light field image that is closer to human subjective evaluation. Compared to other assessment methods, it achieves higher accuracy on multiple public datasets.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of light field image quality assessment, and in particular to a method and system for assessing the quality of referenceless light field images based on light field decoupling. Background Technology

[0002] Light field images, as a novel form of multimedia, aim to provide end users with a more immersive experience. Unlike traditional two-dimensional images that only contain spatial information, light field images record the direction and intensity of every ray of light in a scene, thus containing rich spatial and angular information. This has attracted widespread attention and numerous applications of light field technology in various fields, such as rendering new views, depth estimation, refocusing, and 3D modeling. Research on light field image quality evaluation methods can provide valuable guidance for a range of light field image processing tasks, including the acquisition, compression, and reconstruction of light field images.

[0003] However, the current standardized quality assessment system for light field images still has shortcomings. Although the assessment index based on the human visual system (HVS) is still the most effective method, its complex process and the large amount of human and material resources required make it impractical for real-world applications. Therefore, research on objective light field image quality assessment methods that conform to the subjective perception of the human eye is particularly important. Among them, the objective assessment method without reference has become the main research direction because it does not require the original image information and has a wide range of applications.

[0004] Therefore, the purpose of this application is to propose a referenceless light field image quality evaluation method and system based on light field decoupling to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned technical problems, the present invention aims to provide a device for recycling retired lithium iron phosphate batteries. This objective can be achieved through the following technical solutions:

[0006] This application provides a no-reference light field image quality assessment method based on light field decoupling, which includes using an image patch selection strategy to filter high-quality light field data information on a light field image dataset and constructing its own sample dataset;

[0007] By utilizing the concept of decoupling, the light field image data in the constructed light field dataset is preprocessed to obtain different types of low-dimensional input data for neural networks;

[0008] Using a four-stream convolutional neural network, spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions are extracted.

[0009] By integrating four types of feature information—spatial feature information, angular feature information, and spatial angular consistency feature information in both horizontal and vertical directions—the quality score of the final sample is predicted.

[0010] Specifically, the step of using an image patch selection strategy to filter high-quality light field data information on the light field image dataset and constructing its own sample dataset includes:

[0011] Using edge detection algorithms, image patches with rich structural information in the light field image are selected;

[0012] The selected image patches are arranged into an image patch array, i.e., a mini sub-aperture image array, to construct a new sample dataset;

[0013] Specifically, the step of using an edge detection algorithm to filter image patches with rich structural information in the light field image includes:

[0014] The sub-aperture image at the center of the original light field sub-aperture image array is selected, and only the Y component in its YCbCr color space is retained as the template for the selection strategy.

[0015] The gradient magnitude map of the template image is calculated using the Canny edge detection operator, and its average gradient magnitude is set as the threshold t.

[0016] Using a sliding window of fixed size 32×32, the gradient magnitude map of the entire image template is traversed to filter out image blocks with an average gradient magnitude greater than the threshold t.

[0017] Specifically, the step of assembling selected image patches into an image patch array to construct a new sample dataset includes:

[0018] In the original 9×9 light field sub-aperture image array, select sub-aperture images in odd-numbered rows and columns, and retain only the Y component in their YCbCr color space as the base;

[0019] Image blocks at the same location in each sub-aperture image in the substrate are selected and arranged into an image block array according to the relative positions between the original sub-aperture images. This array is then used as new sample data.

[0020] Specifically, the decoupling concept is used to preprocess the light field image data in the constructed light field dataset to obtain different types of low-dimensional input data for the neural network, including:

[0021] The sub-aperture images in the mini sub-aperture image array are stitched together into a two-dimensional matrix according to their relative positions to generate the mini sub-aperture image matrix, which is used as input one;

[0022] The pixels at the same position in each sub-aperture image in the mini sub-aperture image array are grouped together to generate a mini macro-pixel image matrix, which is used as input two;

[0023] The pixels of each row of sub-aperture images in the mini sub-aperture image array are stacked vertically to generate a mini horizontal polar plane image matrix, which is used as input three;

[0024] The pixels of each column of sub-aperture images in the mini sub-aperture image array are stacked horizontally to generate a mini vertical polar plane image matrix, which is used as input four.

[0025] Specifically, the extraction of spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions using a multi-stream convolutional neural network includes:

[0026] A four-stream convolutional neural network is used to extract corresponding feature information for different types of inputs;

[0027] In the feature extraction process, different network streams use the same backbone network architecture, so that the network has the same ability to extract different feature information.

[0028] Specifically, the process of fusing multiple feature information to predict the final quality score of the selected image patch and calculating the quality score of the original light field image includes:

[0029] The four-stream feature information extracted by the convolutional neural network is concatenated along the channel dimension.

[0030] The concatenated feature information is then subjected to further convolution operations, a global average pooling layer, and a fully connected layer to output the final quality prediction value.

[0031] Specifically, including:

[0032] A sample construction module performs image patch selection on the light field image and reconstructs the sample dataset;

[0033] The light field decoupling module performs dimensionality reduction on the samples in the sample dataset to obtain different types of light field information as input data.

[0034] The feature extraction module utilizes a four-stream neural network to simultaneously extract spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions from the light field data.

[0035] The score prediction module integrates extracted feature information of different types to predict the final quality score.

[0036] In addition, the present invention also provides a referenceless light field image quality assessment system based on light field decoupling, wherein the sample construction module includes:

[0037] The detection unit searches for image patches with rich structural information in the light field image;

[0038] The reconstruction unit reassembles image patches rich in structural information into a mini sub-aperture image array to construct a new sample dataset.

[0039] The optical field decoupling module includes:

[0040] The mini sub-aperture image array is transformed into four low-dimensional data formats: mini sub-aperture image matrix, mini macropixel image matrix, mini horizontal polar plane image matrix, and mini vertical polar plane image matrix.

[0041] The feature extraction module includes:

[0042] A four-stream convolutional neural network is used to extract corresponding spatial feature information, angular feature information, and spatial angular consistency feature information in the horizontal and vertical directions for different types of inputs.

[0043] The score prediction module includes:

[0044] The feature fusion unit connects the four types of feature information extracted by the four-stream neural network in the channel dimension and fuses the feature information using convolution operations;

[0045] The score prediction unit outputs the final quality prediction value by passing the fused feature information through a global average pooling layer and a fully connected layer.

[0046] In summary, this application includes the following beneficial technical effects:

[0047] By employing an image patch selection strategy, the original light field data can be filtered and reconstructed, creating a high-quality sample dataset and solving the problem of preparing high-quality training samples for model training. The use of optical decoupling operations enables the dimensionality reduction of high-dimensional light field data into four formats of low-dimensional data, addressing the issue of how to convert high-dimensional light field data into model input without losing original information. Furthermore, the use of a four-stream convolutional neural network extracts different feature information from different types of input data, solving the problem of how the neural network model extracts and fuses multiple types of feature information from light field data and makes score predictions. In summary, this invention, through the combination of the above technologies, can achieve a highly accurate objective evaluation of light field image quality. Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 This is a schematic diagram of the mini sub-aperture image array of the present invention;

[0050] Figure 3This is a schematic diagram of the mini image matrix of the present invention;

[0051] Figure 4 This is a schematic diagram of the four-stream convolutional neural network of the present invention;

[0052] Figure 5 This is a schematic diagram of the system module connections of the present invention. Detailed Implementation

[0053] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0054] This invention provides an embodiment, such as Figure 1 As shown, a referenceless light field image quality assessment method based on light field decoupling includes:

[0055] S01 uses an image patch selection strategy to filter high-quality light field data on the light field image dataset and constructs its own sample dataset.

[0056] S02 performs a decoupling operation on the light field image data in the sample dataset to obtain different types of low-dimensional input data for neural networks;

[0057] S03 uses a four-stream convolutional neural network to extract spatial feature information, angular feature information, and spatial angular consistency feature information in the horizontal and vertical directions from different network streams for different input data.

[0058] S04 integrates the four feature information extracted above to predict the quality score of the final sample.

[0059] Based on the above embodiments, the present invention provides a preferred embodiment, which is a referenceless light field image quality evaluation method based on light field decoupling, specifically including:

[0060] S11 uses an image patch selection strategy to filter out image patches with rich edge information from the original light field data and reassembles them into a mini sub-aperture image array as new sample data, thereby constructing a sample dataset.

[0061] S12 decouples the light field image data in the sample dataset, transforming the high-dimensional light field data into four types of low-dimensional neural network inputs: mini sub-aperture image matrix, mini macro-pixel image matrix, mini horizontal polar plane image matrix, and mini vertical polar plane image matrix.

[0062] S13 utilizes a four-stream convolutional neural network, employing the same backbone network across different network streams, to extract four types of feature information from the light field image: spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions.

[0063] S14 concatenates the extracted features of different types along the channel dimension, and outputs the final quality prediction value after further convolution operations, global average pooling layers, and fully connected layers.

[0064] In a preferred embodiment of the invention, S11 is implemented as follows: Figure 2 As shown, specifically, in the Win5-LID, NBU-LF1.0, and SHU public light field datasets, sub-aperture images with an odd number of rows and columns are selected from a 9×9 sub-aperture image array to form a new array. This reduces computational cost while increasing the parallax between sub-aperture images and retaining more angular feature information. Considering that the human eye is sensitive to the luminance component Y in the YCbCr color space, the selected sub-aperture images are further processed to retain only their Y color component as the base. Considering that the human eye is sensitive to edge information and that the impact of distortion on light field image data is concentrated at the edges, edge information is selected as the image block selection criterion. First, the gradient magnitude map G of the central sub-aperture image in the sub-aperture image array is calculated using the Canny edge detection operator, and the average gradient magnitude of the entire image is calculated as the selection threshold t, which can be calculated by the following formula:

[0065]

[0066] Where V and H are the spatial resolutions of the sub-aperture images, and G... ij Let the pixel intensity of the gradient magnitude map G at coordinates (i,j) be defined. Using a sliding window of fixed size of 32×32, find image patches on the gradient magnitude map G whose average gradient magnitude is greater than the threshold t and record the position of the image patch. Extract image patches of the same size from the same position in each sub-aperture image in the substrate, and reassemble them according to the relative position between the atomic aperture images to form a mini sub-aperture image array with a spatial resolution of 5×5 and an angular resolution of 32×32, which serves as a new sample. Construct the sample dataset in this way.

[0067] In a preferred embodiment of the invention, S12 is implemented as follows: Figure 3 As shown, specifically, the data format of the sample data is converted. Let the dimensions of the sample light field image be (u,v,s,t), where (u,v) are angular coordinates and (s,t) are spatial coordinates.

[0068] The original sample data, i.e. the mini sub-aperture image array, has a dimension of (5,5,32,32). By writing the 32×32 mini sub-aperture images into a 160×160 empty matrix according to their relative positions on the angular coordinates, the dimensionality is reduced to obtain the mini sub-aperture image matrix, which serves as input one.

[0069] Pixels at the same position in each mini-aperture image in the sample data are grouped into 5×5 macro-pixel blocks. The blocks are then written into a 160×160 empty matrix according to their relative positions in spatial coordinates for dimensionality reduction. This results in the mini-macro-pixel image matrix, which serves as input two.

[0070] By fixing the v and t coordinates in the sample light field data, a set of pixel rows with horizontal parallax is obtained. These rows are stacked vertically to obtain a single horizontal epipolar image with dimensions (5,32). All horizontal epipolar images are generated by traversing v and t, forming a 160×160 mini horizontal epipolar image matrix, which serves as input three.

[0071] By fixing the u and s coordinates in the sample light field data, a set of pixel columns with vertical parallax is obtained. After horizontally stacking and transposing them, a single vertical epipolar image with dimensions (5,32) is obtained. All vertical epipolar images are generated by traversing u and s, forming a 160×160 mini vertical epipolar image matrix, which serves as input four.

[0072] In a preferred embodiment of the invention, S13 is implemented as follows: Figure 4As shown, a four-stream convolutional neural network is designed to extract corresponding feature information for different input data in different network streams. The same backbone network structure is used in different network streams to ensure that the model has the same ability to extract different types of feature information, thereby better analyzing the types of feature information that have a greater impact on the quality of the light field. This can reduce the impact of network hyperparameter settings on feature fusion and thus better analyze the types of feature information that have a greater impact on the quality of the light field. The backbone network contains a single convolutional module and three residual modules. Each residual module contains two convolutional modules. Each convolutional module contains a convolutional layer, a batch normalization layer and a ReLU activation function. In the residual module, the input of the first convolutional module is added before the ReLU activation function of the second convolutional module to improve the model's learning ability. In the convolutional layers of the 1st, 4th, and 6th convolutional modules, a convolution operation with a kernel size of 3, a stride of 2, and a padding parameter of 1 is used. After this operation, the feature map size is halved, and the number of channels in the feature map is changed. The convolutional layers of the remaining convolutional modules all use a convolution operation with a kernel size of 3, a stride of 1, and a padding parameter of 1. These operations do not change the size of the feature map or the number of channels. After feature extraction by the backbone network, the feature information of each network stream has the same feature map size (20, 20) and the same number of channels (256).

[0073] In a preferred embodiment of the present invention, in implementation S14, the four types of feature information extracted by the four-stream convolutional neural network are concatenated along the channel dimension to generate a concatenated feature information of size (20,20) and number of channels 1024. The feature information is then fused through two convolutional modules. Each convolutional module contains a convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional layer uses a convolutional operation with a kernel size of 3, a stride of 1, and a padding parameter of 1, and increases the number of channels of the feature information to obtain a fused feature information map of size (20,20) and number of channels 2048. Then, the feature map is transformed into a feature vector of length 2048 through a global average pooling and flattening layer. Subsequently, the feature vector length is reduced from 2048 to 512, then to 128, and finally to 1 through three fully connected layers, outputting the final quality prediction value.

[0074] Based on the same inventive concept, embodiments of the present invention also provide a referenceless light field image quality assessment system based on light field decoupling, such as... Figure 5As shown, it includes: a sample construction module, a light field decoupling module, a feature extraction module, and a score prediction module. The sample construction module selects image patches from the light field image to construct a sample dataset. The light field decoupling module reduces the dimensionality of the sample data to obtain different types of light field information as input data. The feature extraction module uses a four-stream neural network to simultaneously extract spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions of the light field data. The score prediction module fuses the extracted different types of feature information to predict the final quality score.

[0075] In a preferred embodiment, the sample construction module includes a detection unit and a reconstruction unit, wherein the detection unit uses an image patch selection strategy to find image patches with rich structural information in the light field image; and the reconstruction unit reassembles the image patches with rich structural information into a mini sub-aperture image array to construct a new sample dataset.

[0076] The light field decoupling module transforms the sample data into four low-dimensional data formats: mini sub-aperture image matrix, mini macro-pixel image matrix, mini horizontal polar plane image matrix, and mini vertical polar plane image matrix.

[0077] The feature extraction module uses a four-stream convolutional neural network to extract corresponding spatial feature information, angular feature information, and spatial angular consistency feature information in the horizontal and vertical directions for different types of inputs.

[0078] The score prediction module includes a feature fusion unit and a score prediction unit. The feature fusion unit connects the four types of feature information extracted by the four-stream neural network in the channel dimension and fuses the feature information using convolution operations. The score prediction unit outputs the final quality prediction value by passing the fused feature information through a global average pooling layer and a fully connected layer.

[0079] The following experiments were conducted on three public light field datasets: Win5-LID, NBU-LF1.0, and SHU, to evaluate the model proposed in this invention. The experimental environment was a Win11 system and a Python 3.8 experimental platform. The evaluation metrics were SROCC (Spearman rank-order correlation coefficient), PLCC (Pearson linear correlation coefficient), and RMSE (Root mean squared error). The closer SROCC and PLCC are to 1, and the closer RMSE is to 0, the better the model performance. This embodiment adopted a K-fold cross-validation strategy, where K is 10, and a total of 45 rounds of cross-validation were performed. In each round, 80% of the sample dataset was selected as training samples, and the remaining 20% ​​was selected as validation samples. The median of the metric data after all rounds of training was selected as the final evaluation performance of the model. In the calculation of RMSE and PLCC, the predicted values ​​of the objective method need to be nonlinearly regressed using the logistic function, as shown below:

[0080]

[0081] Where q is the predicted score output by the model, q' is the corresponding nonlinear fitting result for q, and β i (i = 1, 2, ..., 5) are the fitting parameters.

[0082] Table 1 compares the method of this invention with other full-reference / no-reference quality assessment methods on three publicly available light field datasets: Win5-LID, NBU-LF1.0, and SHU. The method of this invention outperforms existing methods in all metrics.

[0083] Table 1

[0084]

[0085]

[0086] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A no-reference light field image quality assessment method based on light field decoupling, characterized in that, include: A sample dataset is constructed by using an image patch selection strategy to filter high-quality light field data on a light field image dataset. By utilizing the concept of decoupling, the light field image data in the constructed light field dataset is preprocessed to obtain different types of low-dimensional input data for neural networks; Using a four-stream convolutional neural network, spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions are extracted. The process of using an image patch selection strategy to filter high-quality light field data on a light field image dataset and constructing a sample dataset includes: Using edge detection algorithms, image patches with rich structural information in the light field image are selected; The selected image patches are arranged into an image patch array, i.e., a mini sub-aperture image array, to construct a new sample dataset; The method utilizes decoupling to preprocess the light field image data in the constructed light field dataset to obtain different types of low-dimensional input data for neural networks, including: The sub-aperture images in the mini sub-aperture image array are stitched together into a two-dimensional matrix according to their relative positions to generate the mini sub-aperture image matrix, which is used as input one; The pixels at the same position in each sub-aperture image in the mini sub-aperture image array are grouped together to generate a mini macro-pixel image matrix, which is used as input two; The pixels of each row of sub-aperture images in the mini sub-aperture image array are stacked vertically to generate a mini horizontal polar plane image matrix, which is used as input three; the pixels of each column of sub-aperture images in the mini sub-aperture image array are stacked horizontally to generate a mini vertical polar plane image matrix, which is used as input four. By integrating four types of feature information—spatial feature information, angular feature information, and spatial angular consistency feature information in both horizontal and vertical directions—the quality score of the final sample is predicted.

2. The referenceless light field image quality assessment method based on light field decoupling according to claim 1, characterized in that, The method of using edge detection algorithms to filter image patches with rich structural information in the light field image includes: The sub-aperture image at the center of the original light field sub-aperture image array is selected, and only the Y component in its YCbCr color space is retained as the template for the selection strategy. The gradient magnitude map of the template image is calculated using the Canny edge detection operator, and its average gradient magnitude is set as the threshold t. Using a sliding window of fixed size 32×32, the gradient magnitude map of the entire image template is traversed to filter out image blocks with an average gradient magnitude greater than the threshold t.

3. The referenceless light field image quality evaluation method based on light field decoupling according to claim 2, characterized in that, The step of assembling selected image patches into an image patch array to construct a new sample dataset includes: In the original 9×9 light field sub-aperture image array, select sub-aperture images with odd number of rows and columns, and retain only the Y component in the YCbCr color space as the base; Image blocks at the same location in each sub-aperture image in the substrate are selected and arranged into an image block array according to the relative positions between the original sub-aperture images. This array is then used as new sample data.

4. The referenceless light field image quality assessment method based on light field decoupling according to claim 1, characterized in that, The method of using a multi-stream convolutional neural network to extract spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions includes: A four-stream convolutional neural network is used to extract corresponding feature information for different types of inputs; In the feature extraction process, different network streams use the same backbone network architecture, so that the network has the same ability to extract different feature information.

5. The referenceless light field image quality assessment method based on light field decoupling according to claim 1, characterized in that, By fusing multiple feature information, the final quality score of the selected image patch is predicted, and the quality score of the original light field image is calculated, including: The four-stream feature information extracted by the convolutional neural network is concatenated along the channel dimension. The concatenated feature information is then subjected to further convolutional operations, global average pooling layers, and fully connected layers to output the final quality prediction value.

6. A referenceless light field image quality assessment system based on light field decoupling, characterized in that, include: A sample construction module performs image patch selection on the light field image and reconstructs the sample dataset; The light field decoupling module performs dimensionality reduction on the samples in the sample dataset to obtain different types of light field information as input data. The feature extraction module utilizes a four-stream neural network to simultaneously extract spatial feature information, angular feature information, and spatial angular consistency feature information in both the horizontal and vertical directions from the light field data. A score prediction module, which integrates extracted feature information of different types to predict the final quality score; The sample construction module includes: The detection unit searches for image patches with rich structural information in the light field image; The reconstruction unit reassembles image patches rich in structural information into a mini sub-aperture image array to construct a new sample dataset. The optical field decoupling module includes: The mini sub-aperture image array is transformed into four low-dimensional data formats: mini sub-aperture image matrix, mini macropixel image matrix, mini horizontal polar plane image matrix, and mini vertical polar plane image matrix. The feature extraction module includes: A four-stream convolutional neural network is used to extract corresponding spatial feature information, angular feature information, and spatial angular consistency feature information in the horizontal and vertical directions for different types of inputs. The score prediction module includes: The feature fusion unit connects the four types of feature information extracted by the four-stream neural network in the channel dimension and fuses the feature information using convolution operations; The score prediction unit outputs the final quality prediction value by passing the fused feature information through a global average pooling layer and a fully connected layer.